Qwen 72B — Hardware Requirements & GPU Compatibility
Chat
Qwen-72B is Alibaba's 72-billion-parameter dense base language model from the original Qwen series, pretrained on over 3 trillion tokens of web text, books, and code. This is the pretrained base checkpoint, not a chat model; Alibaba released it for further fine-tuning and separately published Qwen-72B-Chat as the instruction-tuned counterpart. At this scale, local inference realistically needs a high-end multi-GPU workstation or heavy quantization onto a single very high-VRAM card. The model supports a 32,768 token context window. It is released under Alibaba's Tongyi Qianwen License Agreement, a custom license rather than a standard open-source one, and was published in November 2023. It uses an unusually large, roughly 150,000-token vocabulary for multilingual coverage, and predates and was superseded by the Qwen1.5 and later Qwen2/2.5 generations.
Doesn't fit a 24 GB GPU
Smallest consumer fit → Mac Studio M1 Max (32 GB) · IQ2_XXS · 26.6 GB · ~13 tok/s
Macs from 32 GB (IQ2_XXS · ~13 tok/s)
Other
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Qwen 72B
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
ollama run hf.co/mradermacher/Qwen-72B-i1-GGUF:Q4_K_MQ4_K_M needs offload on the RTX 4090 · 28.5 GB short · 16K context
Download: Qwen-72B.i1-Q4_K_M.gguf45.09 GBmradermacher
Other ways
each opens step-by-stepRun in cloud
Fits on A100 80GB (27 GB headroom) · Q4_K_M
- Generation speed
- ~28 tok/s
- generation speed
- Cost per 1M output tokens
- $13.89
- per 1M output tokens
How Much VRAM Does Qwen 72B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 33.8 GB | 44 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Does not fit: needs 38.5 GB | |
| 34.8 GB | 45 GB | 3-bit small — 3-bit small quantization | Does not fit: needs 39.5 GB | |
| 38.8 GB | 48.9 GB | 3-bit medium — 3-bit medium quantization | Does not fit: needs 43.5 GB | |
| 39.8 GB | 49.9 GB | 4-bit · legacy — 4-bit legacy quantization | Does not fit: needs 44.5 GB | |
| 47.7 GB | 57.9 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Does not fit: needs 52.5 GB | |
| 56.7 GB | 66.8 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Does not fit: needs 61.4 GB | |
| 65.6 GB | 75.8 GB | 6-bit, very good — 6-bit quantization, very good quality | Does not fit: needs 70.4 GB | |
| 79.5 GB | 89.7 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Does not fit: needs 84.3 GB |
Selected Q4_K_M
4-bit medium quantization52.5GBat 16K context
- Weights
- 46.7 GB
- KV cache
- 5.4 GB
- Overhead
- 0.3 GB
Doesn't fit RTX 4090 (reference) · needs 28.5 GB more
Which Hardware Can Run Qwen 72B?
Q4_K_M · 47.7 GBQwen 72B (Q4_K_M) requires 47.7 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 63+ GB is recommended. No consumer GPU has enough memory. 52 devices with unified memory can run Qwen 72B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Pro M2 Ultra (64 GB).
32–40 GB
Q4_K_M · best fit IQ2_M (31.6 GB)
1 GPU · 3 Mac classes · 1 device family
RTX 5090~43Simulate output on RTX 5090M5 Max36 GB~12Simulate output on Mac Studio M5 Max (36 GB)M532 GB~4Simulate output on MacBook Air 13" M5 (32 GB)Orin~5Simulate output on NVIDIA Jetson AGX Orin 32GB64–80 GB
fits · 11.5 GB free · up to 32K ctx
3 Mac classes · 1 device family
M2 Ultra*64 GB~12Simulate output on Mac Pro M2 Ultra (64 GB)M5 Max64 GB~9Simulate output on Mac Studio M5 Max (64 GB)Orin~3Simulate output on NVIDIA Jetson AGX Orin 64GB
By VRAM tier · typical GPUs and Macs at Q4_K_M
Alternatives to Qwen 72B
Qwen family · 28 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Llama 3.3 70B InstructMeta | 79.7 | 51.3 GB | —needs > 24 GB | same size |
| NVIDIA Nemotron 3.5 Bolt 74B A9B BF16NVIDIA | — | 54.2 GB | —needs > 24 GB | same size |
| CodeLlama 70B Instruct HFMeta | — | 23.2 GB | ~29 | same size |
| StableBeluga2Stability AI | — | 23.5 GB | ~29 | same size |
| Apertus 70B 2509swiss-ai | — | 48 GB | —needs > 24 GB | same size |
| DeepSeek R1 Distill Llama 70BDeepSeek | — | 48 GB | —needs > 24 GB | same size |
Specifications
- Publisher
- Alibaba
- Family
- Qwen
- Parameters
- 72.3B
- Architecture
- QWenLMHeadModel
- Context Length
- 32,768 tokens
- Vocabulary Size
- 152,064
- Release Date
- 2023-11-26
- License
- Other
Frequently Asked Questions
How much VRAM does Qwen 72B need?
Qwen 72B requires 47.7 GB of VRAM at Q4_K_M, or 159.0 GB at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 72.3B × 4.8 bits ÷ 8 = 43.4 GB
KV Cache + Overhead ≈ 4.3 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Qwen 72B?
Yes, at IQ2_XS (23.9 GB) or lower. Higher quantizations like IQ2_S (24.9 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Qwen 72B?
For Qwen 72B, Q4_K_M (47.7 GB) offers the best balance of quality and VRAM usage. Q5_K_S (54.7 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 21.9 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Qwen 72B on a Mac?
Yes — Mac Studio M1 Max (32 GB) and 84 other Macs can run Qwen 72B, the smallest at IQ2_XXS. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 47.7 GB of usable unified memory (RAM minus macOS overhead), from Mac Pro M2 Ultra (64 GB).
Can I run Qwen 72B locally?
Yes — Qwen 72B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Qwen 72B?
At Q4_K_M, Qwen 72B can reach ~101 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 47.7 × 0.65 = ~109 tok/s
Estimated speed at Q4_K_M (47.7 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Qwen 72B?
At Q4_K_M, the download is about 43.37 GB. The full-precision BF16 version is 144.58 GB. The smallest option (IQ2_XXS) is 19.88 GB.
Which GPUs can run Qwen 72B?
No single consumer GPU has enough VRAM to run Qwen 72B at Q4_K_M (47.7 GB). Multi-GPU or professional hardware is required.
Which devices can run Qwen 72B?
52 devices with unified memory can run Qwen 72B at Q4_K_M (47.7 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.